Learning obervables of a multi-scale simulation system of urban traffic

Multi-scale modelling is a powerful approach that has been successfully exploited in the context of simulation of traffic and transportation systems. While the paradigm allows the simulation of large cities in a already efficient fashion, the consideration of detailed environments for a precise simulation of pedestrian traffic can be still a demanding task, especially in iterative approaches for the search of optimal solutions. In this context, the paper proposes the application of a supervised machine learning algorithm to learn the observables of a microscopic model of pedestrian dynamics in the simulated environment. The aim is to generate a simpler model that (i) is able to describe the dynamic travel times of pedestrians in the scenario and (ii) can replace the microscopic model in the iterative search of optimal solutions. After a formal description of the approach, the paper provides preliminary results with its application in benchmark scenarios, aimed at analysing its reliability in controlled conditions.